The conditionality principle is a Fisherian principle of statistical inference that Allan Birnbaum formally defined and studied in an article in the Journal of the American Statistical Association, .
Informally, the conditionality principle can be taken as the claim that
Experiments which were not actually performed are not relevant to any statistical analysisand the implicit admonition that unrealized experiments should be ignored: Not included as part of any calculation or discussion of results.
Together with the sufficiency principle, Birnbaum's version of the principle implies the famous likelihood principle. Although the relevance of the proof to data analysis remains controversial among statisticians, many Bayesians and likelihoodists consider the likelihood principle foundational for statistical inference.
Some statisticians in the mid 20th century had proposed that a valid statistical analysis must include all of the possible experiments which might have been conducted. Perhaps a series of desired experiments that each require some uncertain opportunity in order to carry out. The uncertain factor could be something such as good weather for a timely astronomical observation (the "experiment" being the search of the telescopic image for traces of some type of object), or availability of more data resources, such as the chance of discovery of some new fossil that would provide more evidence to answer a question covered by another paleontological study. Another resource issue might be the need for special access to private data (patients' medical records, for example) from one of several possible institutions, most of which would be expected to refuse permission; the nature of the data that could possibly be provided and the correct statistical model for its analysis would depend of which institution granted access and how it had collected and curated the private data that might become available for a study (technically, in this case the "experiment" has already been conducted by the medical facility, and some other party is analyzing the collected data to answer their own research question).
All these examples illustrate normal issues of how uncontrolled chance determines the nature of the experiment that can actually be conducted. Some analyses of the statistical significance of the outcomes of particular experiments incorporated the consequences such chance events had on the data that was obtained. Many statisticians were uncomfortable with the idea, and tended to tacitly skip seemingly extraneous random effects in their analyses; many scientists and researchers were baffled by a few statisticians' elaborate efforts to consider circumstantial effects in the statistical analysis of their experiments which the researchers considered irrelevant.
A few statisticians in the 1960s and 1970s took the idea even further, and proposed that an experiment could deliberately design-in a random factor, usually by introducing the use of some ancillary statistic,
h ,
The conditionality principle is a formal rejection of the idea that "the road not taken" can possibly be relevant: In effect, it banishes from statistical analysis any consideration of effects from details of designs for experiments that were not conducted, even if they might have been planned or prepared for. The conditionality principle throws out all speculative considerations about what might have happened, and only allows the statistical analysis of the data obtained to include the procedures, circumstances, and details of the particular experiment actually conducted that produced the data actually collected. Experiments merely contemplated and not conducted, or missed opportunities for plans to obtain data, are all irrelevant and statistical calculations that include them are presumptively wrong.
The conditionality principle makes an assertion about a composite experiment,
E ,
Eh ;
h
x
E
h ,
xh
Eh~.
The conditionality principle can be formally stated thus:
Conditionality Principle:
If
E
Eh ,
l( Eh,xh r)
E ,
x
E
xh
Eh
An illustration of the conditionality principle, in a bioinformatics context, is given by .
h
h=1,\ldots , 6~.
E1,\ldots , E6 ,
Say that the die rolls a '3'. In that case, the result observed for
x
x3 ,
E3~.
E1,E2,E4,E5,~~or~~E6
x1,x2,x4,x5,~~or~~x6 ,
x3 ,
E3 ,
x3 ,
E3 ,
The conditionality principle says that all of the details of
E1,E2,E4,E5,~~or~~E6
x3 ,
h
E3 ,
x3 ,
h=3~.